We describe a polyhedral modeller that uses software extended-precisio
n integer arithmetic to guarantee numerical reliability. By careful de
sign, the performance of the modeller is not much different from the p
erformance that a floating-point modeller might have. The modeller per
forms Boolean set operations exactly; to prevent growth of coordinate
bit-length, affine transformations require coordinate rounding and hen
ce are approximate. A new algorithm for reconstructing polyhedral inci
dence information after rounding is given. Copyright (C) 1996 Elsevier
Science Ltd